Case study · public proof pages

A card matcher that refuses to guess.

EngagementA TCG portfolio and market-intelligence platform (client and product names withheld). Fixed-price, three milestones: discovery audit, verified execution, validated queue drain. Database and verification side by Muhannad Ahmed; the platform's own developer owned the product side. The ruleThe platform's strict matcher opens with a warning comment, quoted on every page here: a wrong match becomes a wrong arbitrage becomes the wrong card bought. Return null rather than guess. The work below verifies a matcher built on that rule, without softening its coverage cost. Public renderingThese are the engagement's real deliverables, re-rendered for public reading. Client, product, and niche-game names are withheld; per-card rows, real listings, repo identifiers, and client-code excerpts are removed or generalized. Every aggregate metric is carried unchanged. The M1 and M2 pages are published as summaries because the full documents enumerate the client's private schema. Re-rendered 2026-07-05.

99.40%

precision on the priority stratum, production matcher scored against all 6,190 labeled sample listings

2,481

audited queue-drain commits verified live on their proposed id; a closed cohort, stable across five production reads

foreign-title wrong-card risk on the flagship game, by construction; the ported-guard gap on other games is measured, not hidden

M1 · Discovery: schema audit and migration plan →

A read-only audit of a 10-game, 17.5M-row production database: scale picture, drift cases, dead crons, a freshness gap on a money surface, and a 10-action migration plan where every action ships with a rollback and a test. Published as a summary; the full document enumerates the client's schema.

delivered 2026-05-12 · public summary 02

M2 · Verification against a 6,190-row labeled sample →

The production matcher scored against every labeled listing, not a hand-picked set: 96.59% overall, 99.40% on the priority stratum, disagreements classified row-by-row instead of averaged, a fix's coverage cost measured and reported in the direction the evidence supports, and one draft claim publicly withdrawn after EXPLAIN ANALYZE.

delivered 2026-05-17 · public summary 03

M3 · Delta report: the queue drain, measured →

An audited batch drain of the near-miss review queue: 2,750 proposed, 141 held back as higher-risk, 2,481 verified live on their proposed id across five production reads. Closed-cohort accounting separated from open-population snapshots, plus two honest observations about what the live sync does on its own.

pinned 2026-05-27 · page-level scrub of the delivered report 04

Per-game coverage, runway, and the accuracy gap →

One live read, three readings: coverage spread 94% to 36% across the four ingested games, a 112,693-card expansion runway already catalogued with zero ingest, and the durable foreign-title wrong-card gap that is zero on the flagship game by construction and open everywhere the guards were never ported.

live read 2026-06-09 · page-level scrub of the delivered report 05

Arbitrage read-path: the unbuilt cross-link, sized →

The two halves of the arbitrage trade exist on separate surfaces and nothing joins them. Sized honestly: $56,598 of addressable spread across 252 cards on a roughly 10%-built surface, a floor on a fraction rather than a headline, with the one wrong-card pair found in 232 name-checked buy candidates reported instead of rounded away.

live read 2026-06-09 · page-level scrub of the delivered report

Muhannad Ahmed · database and data-engineering side of the engagement. The matcher, product, and platform are the client's; the audits, verification harnesses, drain accounting, and these reports are the work being shown.